Meta: Large Language Models Will Not Get to Human-Level Intelligence

Core Argument - Meta's Chief AI Scientist Yann LeCun criticizes the current definitions of artificial general intelligence (AGI), stating that scaling text-based large language models (LLMs) alone is insufficient to achieve AGI [1] - LeCun disagrees with OpenAI CEO Sam Altman's claim that AGI has already been achieved and that the focus is now on superintelligence [1] Limitations of LLMs - LLMs, as they exist today, cannot achieve human-level intelligence due to their reliance on autoregressive text completion, which lacks the multimodal understanding of human brains [2] - Current AI systems are mostly "narrow AI," excelling in specific tasks but failing when tasks deviate slightly from their training [2] - Scaling LLMs is reaching a point of diminishing returns, making further improvements expensive and less effective [5] Challenges in AI Development - AI systems struggle with physical tasks, such as plumbing, due to the complexity of understanding and manipulating the physical world [4] - Even with advancements, AI systems are far from matching the physical world understanding of animals like cats or dogs [5] - The cost of scaling LLMs remains high, with OpenAI's ChatGPT Pro not yet profitable despite its $200/month subscription fee [6] Progress in AI-Powered Robotics - Generative world models are emerging as a cost-effective and low-risk solution for training AI-powered robots in virtual environments [6] - Nvidia's Cosmos platform enables developers to create synthetic data for training physical AI systems like robots and autonomous vehicles [7] - Google DeepMind and Fei Fei Li's World Labs are investing in generative world models, with significant funding from prominent Silicon Valley figures [8] - The "ChatGPT moment" for robotics, driven by world models, could be 3-5 years away [8] Future of AI Assistants - AI agents are expected to become more common as people grow accustomed to using specialized AI assistants for specific tasks [9] - These AI assistants will be task-specific and not capable of performing activities from scratch without specific training [9]